Visual SLAM algorithm based on improved YOLOv10 and coordinate attention for dynamic scenes
TL;DR
A visual SLAM algorithm based on an improved YOLOv10 and coordinate attention based on the ORB-SLAM3 framework achieves an improvement rate of 53.56% and 9% in the Absolute Trajectory Error compared with traditional algorithms, realizing the effective elimination of dynamic features and the consistent construction of static environment maps.
Abstract
To improve the localization accuracy of visual SLAM in high-dynamic scenes, this paper proposes a visual SLAM algorithm (OY-SLAM) based on an improved YOLOv10 and coordinate attention. Built on the ORB-SLAM3 framework, the algorithm deploys the improved YOLOv10 network in the front-end thread to realize dynamic feature detection, and then leverages geometric feature constraints to achieve real-time elimination of dynamic features. Comparative experiments are conducted on the TUM and EuRoC datasets, and the results demonstrate that the proposed OY-SLAM can effectively eliminate dynamic features. In particular, in high-dynamic scenes, it achieves an improvement rate of 53.56% and 9% in the Absolute Trajectory Error compared with traditional algorithms (ORB-SLAM2 and VDO-SLAM), realizing the effective elimination of dynamic features and the consistent construction of static environment maps.